Set 16..4096-byte bound on cache-block size

16 B suffices to subsume all aligned accesses (including the Q extension).
Spike does not actually rely on this property, but in some real systems,
it is impractical to guarantee atomicity across cache lines.

4096 B suffices to prevent cache lines from spanning pages (which would
require multiple TLB accesses).  This one is a bug fix, since we were
not performing multiple TLB accesses in this case.